When Teamwork Fails: A Structural Autopsy of Multi-Agent Failures in Multimodal Clinical Risk Prediction

While multi-agent systems (MAS) have driven substantial advancements across various domains, their efficacy in high-stakes clinical risk prediction remains notably limited. This limitation is especially pronounced in multimodal environments, where the integration of fragmented clinical data introduces unexplored architectural vulnerabilities. To investigate this performance gap, we introduce a novel diagnostic framework to conduct a structural autopsy of MAS failure modes, targeting multimodal Intensive Care Unit (ICU) in-hospital mortality prediction. By stress-testing established MAS collaboration frameworks, we reveal that the tested MAS are critically hindered by compounding vulnerabilities. Specifically, we identify three primary structural failures that expose a fundamental flaw: in the zero-shot evaluation paradigm, as agentic complexity increases, MAS develop a disproportionate hyper-sensitivity to input perturbations. We open-source our diagnostic framework to facilitate rigorous failure mode analysis of MAS across other domains.